Causal Context Graph (CCG) connects marketing context, decisions, experiments, and aggregate outcomes so your team can reuse prior learning. It helps marketers, analysts, marketing scientists, and agentic systems reason from evidence instead of treating every historical pattern as a causal result.

Marketing systems are good at storing artifacts: an audience, campaign, journey, metric, or dashboard result. They are less reliable at preserving the central causal question behind those artifacts:

• What business objective did the campaign support?
• Which customer population was eligible?
• What treatment did the control group not receive?
• Which outcome defined success?
• What did the experiment estimate, and with what uncertainty?
• Under what conditions should that learning be reused?

CCG connects those facts into reusable organizational memory. Despite the name, CCG is not an identity graph of individual people. It organizes context about how your company markets to customer populations and what aggregate effects you have measured.

📘

Note

Context without outcomes is description. Outcomes without context are difficult to reuse. CCG connects both.

CCG at a glance

CCG is not one model, score, or database technology. The "graph" is the set of meaningful causal relationships among marketing entities. Its value comes from preserving those relationships when evidence is retrieved and reused.

Why causal context matters

Correlational intelligence looks at what has happened historically. Causal intelligence asks what changed because of an intervention.

For example, assume you test incremental order revenue across two audiences:

• A lower-propensity audience
• A higher-propensity audience

A correlational approach may prioritize the higher-propensity audience because those customers dominate historical order data. A randomized experiment can show a different result: the lower-propensity audience may generate positive incremental revenue, while the higher-propensity audience may show little lift or even negative lift because the campaign cannibalized purchases that would have happened anyway.

CCG keeps that distinction available for future decisions. It helps your team reuse the lesson as causal evidence, not just as another historical correlation.

What the graph represents

The underlying GrowthLoop and measurement systems remain the sources of truth for audiences, journeys, metrics, and results. CCG makes the relationships among those objects usable as evidence.

Evidence hierarchy

CCG combines information with different scientific status. Use the evidence layer to understand what a record can and cannot support.

How CCG is created

CCG follows a continuous evidence cycle.

Collect marketing context

CCG begins with organization-level context and marketing metadata:

• Brand and market description
• Metric definitions
• Audience and journey definitions
• Treatments and allocation
• Campaign and channel information
• Experiment or decisioning configuration

Attach aggregate outcome evidence

Where measurement is available, CCG connects aggregate results to the population, treatment, metric, and decision that produced them.

The graph is designed around aggregate evidence, not customer-level event rows. That distinction reduces data exposure and keeps causal interpretation focused on the experiment or policy instead of an individual profile.

Create concise semantic context

AI-generated descriptions make technical marketing objects easier to retrieve and reason about. Generation is constrained to supplied metadata and results, with instructions not to invent facts or row-level data.

Generated text is a semantic interface to the evidence. It helps agentic systems retrieve and explain relevant context without turning generated language into new statistical evidence.

Reuse evidence in the next decision

Creative and decisioning systems can retrieve relevant organizational evidence before making a recommendation. Evidence from the same organization and metric is preferred over generic world knowledge.

Historical evidence informs a starting point. New live evidence should update and eventually dominate that starting point.

Controlled audience experiments

An audience experiment compares outcomes under a treatment with outcomes under a control condition.

Let:

• Y(1) be the outcome a customer would have under treatment.
• Y(0) be the outcome the same customer would have under control.
• T identify the assigned condition.

The individual causal effect, Y(1) - Y(0), is never directly observable because each customer experiences only one condition. Random assignment makes the groups comparable in expectation, allowing the average difference between groups to estimate the treatment effect for the experimental population:

Estimated effect = average outcome in treatment - average outcome in control

Worked example

Suppose a randomized win-back experiment reports:

Then:

• Absolute lift = 4.8% - 4.0% = 0.8 percentage points
• Relative lift = 0.8% / 4.0% = 20%
• Cumulative impact = 0.008 x 100,000 = 800 incremental purchases

A possible semantic representation in CCG is:

In a randomized test among the eligible win-back audience, the treatment increased 30-day purchase rate by 0.8 percentage points, or 20% relative to control. Applied to the observed treatment group, that difference represents approximately 800 incremental purchases.

AI Decisioning evidence

GrowthLoop AI Decisioning uses Bayesian contextual multi-armed bandits to adapt decisions as evidence accumulates.

CCG can preserve:

• Candidate arms and their marketing meaning
• The optimization metric
• Arm-level reward and population summaries
• Customer-context features used by the policy
• Current parameter estimates and uncertainty
• Observed differences across customer segments

Adaptive evidence is valuable, but it requires different interpretation from a fixed randomized experiment:

• Allocation probabilities can change over time.
• Customers with different contexts can receive different actions.
• The observed sample for an arm is partly selected by the policy.
• Naive arm averages may not estimate a population-wide treatment effect.
• Valid offline comparison can require propensity weighting or another policy-aware estimator.

How historical evidence informs a new decision

CCG supports an evidence hierarchy for cold starts:

1. Prior experiments from the same organization using the same metric.
2. Prior experiments from the same organization using a closely related metric.
3. Organization, brand, audience, channel, and campaign context.
4. General marketing knowledge.
5. A conservative fallback when evidence is weak.

This helps decisioning and agentic workflows start from relevant organizational memory while keeping room for exploration, live learning, and human review.

Marketing use cases

CCG can support several decisioning and agentic workflows:

• AI Decisioning can use CCG as an evidence-informed starting point before a new decision has accumulated enough live observations. Relevant same-organization experiments, metric history, audience context, and journey context can initialize expectations and uncertainty for candidate actions.
• Creative agents can use CCG to generate content variations grounded in uploaded seed content and prior performance context.
• Agentic audience workflows can use connected context to help create outcomes-optimized audiences. Marketers can relate an audience hypothesis to the intended outcome metric and to learnings from comparable experiments.

Privacy, isolation, and governance

CCG is designed around organization-scoped marketing context and aggregate evidence.

Current boundaries

Do not treat CCG as:

Generated context can also become stale or contain summarization errors. Important decisions should retain traceability to the original artifact and measurement.

Glossary

---

Next, learn how to enable AI Studio for a dataset group, then use AI Studio to start building with organizational context.